Leonard Pickard
This channel features a range of podcasts exploring the frontiers of biotechnology, psychedelics, artificial intelligence, creativity, cognitive enhancement, and more.
We interview CEOs, researchers, philanthropists, and public intellectuals who are shaping the questions—and sometimes the controversies—about what may come.
Here, you’ll encounter talented, passionate, and visionary individuals.
Join us to share in the green-growing edge of life and courageous thought, as we explore the potential futures of our species.
William Leonard Pickard
Leonard Pickard
Internet Creator Vint Cerf on AI, AGI & What Comes Next
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
🎙️ JLS Podcast: Hosted by William Leonard Pickard
🚀 About Vint Cerf
Vint Cerf is widely recognized as one of the architects of the modern internet. As a key designer of the TCP/IP protocols that allowed computers to communicate globally, Cerf helped lay the foundation for the digital world. A former DARPA program manager and longtime leader at Google, he has spent decades thinking not only about technology itself, but about its impact on civilization, knowledge, and the future of humanity.
🌐 From the Birth of the Internet to the Age of AI
Cerf reflects on the early days of networking—from ARPANET experiments and the first commercial email systems to the explosive rise of the public internet. What began as an academic and government research tool became a planetary nervous system, reshaping communication, commerce, and culture in ways few could have predicted.
🤖 Artificial Intelligence: Powerful, Imperfect, and Already Here
When asked whether AGI exists, Cerf offers a nuanced answer: in some domains, yes. Modern AI systems can outperform humans in recall, pattern recognition, and certain specialized tasks. Yet they still hallucinate, make mistakes, and lack deeper grounding. He sees current AI not as magic consciousness, but as astonishingly capable systems that continue to evolve through specialized models and layered intelligence.
🧠 Meaning, Semantics, and How Machines “Think”
One of Cerf’s most fascinating insights centers on language models as engines of meaning. Rather than merely predicting words, these systems operate through embeddings, relationships, and semantic structures. While different from the human brain, he suggests they are processing representations of meaning in ways that are surprisingly effective—and deeply worthy of study.
⚠️ AI Agents, Risk, and the Need for Guardrails
Cerf expresses particular concern about autonomous AI agents that can act in the real world—handling finances, infrastructure, or decision-making without sufficient oversight. He emphasizes the need for audit trails, constraints, accountability, and thoughtful design before handing powerful systems the keys to human institutions.
🔬 AI and the Solving of Great Scientific Mysteries
From protein folding to medicine, physics, and chemistry, Cerf believes AI may help solve problems humans have struggled with for generations. Yet he cautions that discovery requires more than analyzing known data—it may also demand new experiments, new sensors, and theories beyond our current understanding. AI may become a partner in discovery, but not the whole story.
⚛️ Quantum Computing and the Next Frontier
Cerf discusses quantum computing with both excitement and realism. While quantum machines may revolutionize specific classes of problems, they are not universal magic boxes. Challenges like coherence, scaling qubits, and quantum networking remain immense. Still, he sees enormous potential in combining future computational tools with scientific exploration.
📱 Technology, Dependence, and Human Fragility
Despite his optimism, Cerf warns that society has become deeply dependent on digital tools. Phones, authentication systems, banking, communication, and daily logistics now rest on fragile technological layers. If those systems fail, modern life can quickly unravel. Progress brings power—but also vulnerability.
🌌 Wonder, Discovery, and the Unknown Universe
Cerf remains driven by awe. He speaks of black holes, gravitational waves, deep time, and mysteries physics has yet to solve. For him, the greatest excitement lies not in what we know, but in what we have not yet learned. Human tools—from telescopes to AI—may reveal realities still hidden from us.
🏆 Final Takeaways
A profound conversation with one of the builders of the digital age. Vint Cerf offers a rare blend of technical wisdom
Dear friends, today's conversation is with Dr. Vince Zurf, one of the principal architects of the Internet. What you're about to hear is not a biography and not a technical lecture. It's a conversation with someone who helped design the infrastructure of modern civilization. And who is now watching artificial intelligence arise as a second-grade transformation? We move quickly. Across the decades, across individuals, even across the planet. It feels a little dense. Stay with it. This is a rare opportunity to hear how someone who shaped the first digital revolution is thinking about the next. It was fundamental in the development of how computers talk to each other. And a funny little anecdote. When we first became acquainted, it turned out that we both had had a hamburger and a little biker bar outside of Stanford called Rosati's. It's a lovely place. It's still there. There's a plaque at Rosati's, it says in 1971, the first email message was uh sent from the beer garden in back of this place. And uh can you speak to that email message event?
SPEAKER_00Okay, so I need to make several uh amendments here. First of all, I was just a program manager and eventually one of the principal uh program managers. Uh the the first plaque of importance in that area related to internet is uh is in the Cabana Hotel on um El Camino Real. El Camino Real. And in that hotel in 1973, Bob Kahn and I spent two days writing our first formal paper on the design of the internet. So there's a plaque in the hotel next to the elevators saying in the you know summer or or fall of 1973, Benserf and Bob Kahn, you know, in uh invented the internet. So that's one plaque. Now, the plaque that you're talking about is a 1976 uh remembrance. That was the first time that the packet radio network in the Bay Area, which was run by SRI International and sponsored by DARPA, was used to go through a gateway into the ARPA net. So it was the first two network interconnection of disparate packet switch nets. The SRI team decided to demonstrate that it could communicate from Rosades through packet radio through a biker from the biker bar. Yeah. So and that's well documented uh in a report from uh their report was actually composed and sent from Rosadis uh to uh DARPA uh from the terminal unit that was uh in use in the packet radio van.
SPEAKER_01I'll never have a hamburger there again without thinking of DARPA. So you watch the internet uh move from academic network to global nervous system. Uh does AI feel like a continuation of that arc or a rupture?
SPEAKER_00Well, interestingly enough, it is, but perhaps not quite in the way that you might think. Uh if we go back and look at history uh of the internet, what we discover that some of its creation was a consequence of research in artificial intelligence that the Defense Department was funding in the 1960s. So if you if you just just track uh AI for a second, it starts uh with a speculative meeting at Dartmouth in 1956. John McCarthy and others are at the meeting, and if my recollection, not personal, but uh from reading, is that uh that McCarthy introduced the term artificial intelligence at that meeting. Um by 1961, uh McCarthy has left MIT and gone to Stanford University, and by happenstance, I arrived there as a as an undergraduate that same year. So I am but I am aware of, but not involved in any of the AI research because I'm a math major.
SPEAKER_01Some of the uh cool early email messages that were sent. We know about hello that sort of was truncated to low. Uh, do you recall any of the early messages? Uh were you guys doing arm waving?
SPEAKER_00Ray Tomlinson you know was asked what was the first message you sent, and he says, I can't remember, and it was probably you know just you know, query oop or something. Um because you know nobody thinks of these things as as having uh the massive impact that they have uh at the uh at the beginning. So I honestly don't remember.
SPEAKER_01You're speaking about the impact, the the vast impact of the internet on society. When did you realize the internet had escaped its original academic intent?
SPEAKER_00It took a while. I didn't really sense how significant this was going to be until 1988. The incident that really triggered my uh realization uh was a show, an exhibit called Interop that had been founded in 1986 or so by Dan Lynch. And uh Dan uh started this as an educational thing, just briefing people on how the internet worked and how the protocols work and how you could implement them, but it quickly turned into an exhibition because companies like Cisco wanted to show their stuff. No, so in 88, uh 50,000 people showed up at the Millsconey Center. The then president of 3COM was Eric Benemo. And so we're walking in, and the first thing we see is a three-story Cisco demonstration and just display. So I turned to Eric and I said, Eric, how much does that cost? And he said, $250,000. Well, that was a lot of money in 1988, I guess it still is. I'm just sitting there with my jaw on the floor thinking, wow, you know, somebody thinks the internet's gonna make money. And part of the reason by that time, we're talking '88 now, uh, is that the large government-sponsored networks, the NSF net from the National Science Foundation, the NASA Science Internet from NASA, and the um ES Net, Energy Sciences Net from DOE, and ARPA, the ARPANET, were only allowed to be used by people who had government research contracts. So I got to thinking, well, how can we break that appropriate use policy and allow commercial traffic to flow on the government-sponsored backbones?
SPEAKER_01A massive concept.
SPEAKER_00So uh I I had worked after I left ARPA, I went to MCI and they built something called MCI Mail, which was a commercial email service. So I asked the Federal Networking Council for permission to connect the commercial MCI mail system to the internet.
SPEAKER_01Oh my god.
SPEAKER_00And you know, as an experiment. Game on. And you know, of course, I was pretty sure I could do it because I'd been involved in both the design of MCI mail and the internet. But I I build it as an experiment. And of course, the uh the subtext was break the rule. So uh for instance, and allow commercial traffic to flow on the government backbone. So they let me do it. And in 89, around mid-year, uh my colleagues at uh CNRI, Bob's company, uh announced that we had an interconnect between MCI Mail and the Internet. So anyone on the internet could talk to anyone on MCI mail. And as soon as we announced this, of course, the other commercial email providers who had their little fiefdoms like telemail or on-time or um I think DICECO had another one. Um said, Well, wait a minute, they can't have that special privilege. We want to connect to the internet too. And so the Federal Networking Council said okay to that. And then they discovered that their little islands were suddenly interconnected through the internet. So anybody on MCI Mail could talk to anybody on time on time, for example.
SPEAKER_01Oh, oh, what an insight!
SPEAKER_00Suddenly they didn't have the protective moat around their services. The second consequence of that is that uh three commercial internet service providers are announced in 1989. The first one was UUNet, uh, which is headquartered here in Virginia. The second one was PSINet, which started out in New York but had headquarters here in Virginia, and the third one was SurfNet out in the uh San Diego area, run by General Atomics. So three commercial internet services start in 1989, and that is a significant beginning. The World Wide Web gets announced in December of 91. The mosaic browser with the graphical user interface is announced and released in 1993. Netscape Communications gets started in 1994. Netscape, yes. Right public in 1995, the stock goes to the roof, and the dot boom is on.
SPEAKER_01Then what outcome of uh of global connectivity uh through the internet has surprised you most?
SPEAKER_00What did you expect? The fact that we uh found it valuable and even necessary to build out a massive uh sub-sea network and an intercontinental and continental network fiber network in order to interconnect our data centers together and to interconnect the data centers to the public internet. Uh the design of the of the uh computing systems at Google uh is designed for resilience. And so whenever you have an interaction with applications in our cloud, we replicate the data in other data centers. So even if we lose a whole data center, nobody loses any data. And in order to do that, we're constantly copying data back and forth among the data centers in order to provide that resilience. Uh and that requires a huge amount of capacity. So a special purpose network was built to accomplish that objective. Then, of course, you have to get to the users. You know, the only way the users could get to our data centers is to go through the public internet. So we had to interconnect another network, a second network we built to connect our data centers to the public internet. And of late, we are combining everything together into one gigantic alpha net uh in order to uh be more efficient in our use of the communication resources. But it was it was necessary to do that much investment in the data centers themselves and in the networking capability. So we're we have a huge capac subseed capacity.
SPEAKER_01And across all the platforms. Uh I every time I come up with Gmail, I see Jim and I looking at me, and it's sort of uh easy to use, and we go right to it. Uh do you think uh all this connectivity increases understanding?
SPEAKER_00No, uh it doesn't increase it necessarily, but it creates a potential for understanding to increase. And the reason is the basic reason that internet was important. There's the potential for finding things, the shareable things, anywhere in the world if people are willing to make them visible. And so the Google search engine and other search engines help people all around the world find information that they're interested in. The most recent evolution of that, of course, is the large language models. Things like uh ChatGPT and Google Gemini and some of the others have incorporated an enormous amount of information which they have harvested off the internet and make even more accessible. The the interaction is fluid, it is uh cogent, uh, and it is deep.
SPEAKER_01I have to ask you this, of course. Uh are we do we have AGI now?
SPEAKER_00I think the answer is yes, in some domains, uh these systems are as good or better than humans are, uh, certainly at recall. It's it's astonishing how much information is readily available, quickly available, and can be assembled in useful ways through some of these large language models. Um, but we also see that they hallucinate. We also see that they make mistakes. An example, of course, is computation. They were just bad at arithmetic. That's been corrected uh in many ways by having specialized um models that have been trained on particular material and trained to be good at whatever that subject matter is. And the uh the large language models that are foundational are now being uh learning, I'll say it that way, they are learning to call on specialized models which are more efficient and and more functionally powerful to answer specific questions. And so you're getting the smart graduate student, so to speak, who knows who to go to to get answers if he or she doesn't have them uh on their own. And so this evolution continues. Uh the possibility of the system uh asking itself questions and and training itself to go and find resources doesn't seem out of reach. And so a self-learning kind of system is entirely uh imaginable. I won't say feasible because I don't know enough to know how hard or easy it would be to make a system learn to teach itself more, but it doesn't seem out of the out of uh the realm of possibility that you could do that.
SPEAKER_01Do you think that we're at the point where um previously intractable problems in mathematics and physics and chemistry can be resolved, problems that uh maybe at the limit of human comprehension, the limit of human ability? Uh I saw an incident uh some months ago where uh Paul Hurdins, an early uh mathematician, quote a character, uh, put out a set of um of problems that were considered relatively insoluble, and suddenly was announced that these problems had been solved by an advanced model. And it turned out that no, there was no deep reasoning involved, that the model had become very good at searching uh worldwide and came up with an arcane paper or citation that was the solution to the EROS problem sets. Uh the the question then is do you think that uh serious um problems such as the standard model gravitational mission may be solved by uh AI soon?
SPEAKER_00Yes. I mean, I I think it's it's I think Demas Hossipis is also uh, I believe, of that opinion. He runs Google Deep Grand, for example, and uh was instrumental in many of the uh advances that have been made in AI at Google over the past decade or so. Um, I am well, look at Alpha Fold as an example of something which has figured out how uh proteins fold up and what their shapes are, and that in that implies what their functionality could be. It might also uh inform the search for uh drugs to interfere with certain kinds of disease processes. Uh so in certain domains, I would not be at all surprised to find that uh problems get solved using these AI systems that humans haven't been able to uh resolve. And part of your implication of finding out some just the ability to see all of the content that might be relevant uh and to recognize that it might, you know, composing it might be a powerful thing, um, seems to me very um obvious that uh that we will see problems solved of that kind anyway.
SPEAKER_01You know, Dim Demis uh Hosivas um uh uh spoke uh very boldly uh after his Nobel that within the next 10 to 20 years, all uh maladies, all human illnesses would be solvable by advances of medicinal chemistry, AI-driven, hence isomorphic labs of of uh Google's deep mind subdivision.
SPEAKER_00Well, you know, here's here's a place where I might uh argue that that is um it's possible, but whether it's probable is a little uncertain in my mind. The reason for this has to do with uh the depth of our understanding of uh the way the world works. Um these models are capable of ingesting and working with that which we've already discovered. Uh, the problem is that we haven't discovered everything yet. And it's not clear whether these models are ready to do their own experiments, for example, to posit theories and and to conduct experiments to validate the theories. Um look at what happened uh when we discovered that atoms were not uh unitary, that they actually were composed of quarks.
SPEAKER_01And probabilities, electron positions.
SPEAKER_00Fundamental particles were composed of quarks, and atoms, of course, composed of protons and neutrons and electrons. So um we when we when we have learned to measure something that we couldn't sense before, we've learned new science. So it's impossible, it is certainly possible that an AI system could design an experiment to measure or detect something that we have not been able to do before. Think about LIGO. Uh we were finally able to test the theory that Einstein posited 100 years ago that uh the collision of two black holes would create waves in the in space-time, gravitational waves that we could detect, except that they were so minute when you're a billion light years away that you had to have an incredibly sensitive detector to do that. Well, we have several.
SPEAKER_01We do now.
SPEAKER_00But we wouldn't even have we speculated about that based on his theory, but we couldn't validate it. So I think AI is going to help us design new experiments. It whether it will help us posit new theories, I'm I'm unsure.
SPEAKER_01Ben, what do you what do you think about uh the emergent properties of AI and deep neural networks, the black box where we really don't understand how some of these solutions uh uh are arrived at uh uh within the black box. Uh and then the the emergent properties like uh Lee Settle's uh Go game, a move 37, which had not been anticipated by Go masters in 5,000 years, suddenly the machine makes the move, and all the GoMasters of the world say that's a mistake. The machine made a mistake, but down the line, it was the critical uh solution, the move that changed everything. Um as an emergent property of generative adversarial networks, machines training machines. Uh, do you do you see emergent properties coming out that we don't anticipate, that we don't program for? Uh is there a a striking novelty to that? Uh, do you worry about it, worry about it? Let me parse this a little bit.
SPEAKER_00Um, an example of an emergent property that I am terrified by is agents and agents that take actions that have real-world consequences. If we're not very careful about how those are built, uh they may go off and do things that you didn't expect or even want, you know, like managing your finances, or you say, I need transportation, and it goes off and buys a Ferrari that you can't afford. Who knows? So agent interactions that are not constrained and uh and are not logged. I mean, we need audit trails for these kinds of automated processes. So if they do go wrong, and some will, how do we find out what happened? You know, who done it?
SPEAKER_02Who did it?
SPEAKER_00And then there are all kinds of questions about liability and insurance and so on and so forth. So agent technology uh looms uh large in my mind as a an area where uh we need to think very carefully about how to manage the process. That is that's an one example of what is clearly emerging right now.
SPEAKER_01Do you think uh there could be malevolent um agent swarms done by say hostile uh actors that uh could collapse financial systems or electrical grids?
SPEAKER_00Uh well, uh certainly. I mean, we know that there our existing infrastructure has vulnerabilities. We experience them all the time. You get a snowstorm, you lose power. It takes a while to put it all back together again. Yeah, there are malware attacks uh like colonial pipeline that That could shut things down. So we should be worried about that because if we turn things over to uh to AI-based agents, uh, but we haven't thought our way through uh how to make sure they understand the consequences of their actions or can detect the consequences of their actions, they may do things that uh um create unwanted and unexpected outcomes. So we have to think very carefully about giving agents the ability to interact with the real world. Um I am I I was thinking very hard about the AGI question uh this morning, and one of the things that um became very apparent is that the language models are dealing fundamentally with semantics. And as you start to learn more and more about how these black boxes are built, uh you begin to appreciate the um level of semantic uh element that is inside the systems. Uh, you probably know that uh when text is ingested into the system, it gets tokenized. And a token could be just a word, it could be a phrase, or it might even be a part of a word, but but these semantic units get labels, they they get unique labels. And so there are lots and lots of different tokens that are part of the system. Then the question is well, what is the meaning of that token? And there is an embedding which takes place. So when you create a token, the next question is, what uh semantics should I associate with that token? Well, imagine that you have 2,000 different uh terms that uh semantic terms or that uh that you can use to describe what that token means. So you're embedding the token in a 2,000-dimensional space. What and then while the training goes on, when you're trying to get the large language model neurons to be adjusted to produce outcomes.
SPEAKER_01I mean weights to the uh layers of the neural network, this sort of thing.
SPEAKER_00It's the weights in the neurons. Uh, I mean, there are degrees of freedom in this system that are striking. You know, there are multiple layers, the layers can be structured differently. The way the information is passed from one layer to another, the way the weights are passed from one layer to another can change and vary, convolutional or direct pass or restrictions and so on. But then there's something else that's very important that I didn't appreciate until I started spending my weekends with Jim and I learning about how AI work. The embedding is also adjusted during the training. And so the token, the semantic value of a token is refined during the process of training the neural network. And so there's there's deep semantics going on here.
SPEAKER_01That's only reasonable. It's a recursive training.
SPEAKER_00And so we we all should appreciate that semantics is what thought is about. That's what thinking is. It's semantics, it's meaning. And so these systems are dealing in a very direct way with a representation of meaning. And the fact that they do so well suggests that the representation which we have artificially created is very effective. It may not be the same as the way our brains work, but it works in ways that are just surprising to us and in many cases extremely beneficial. I mean, think about the vibe coding that's going on today. There's a rapid move towards getting your neural network to do your programming for you.
SPEAKER_01I'm hooked on vibe coding in the last few months, uh Python and this sort of thing. I uh spend hours a day on it, uh, it's uh glorious. What a what an advance. And now now anyone can code. You just uh you get on your appropriate AI model, it's super friendly, and off you go.
SPEAKER_00So this is in in a in a dim way, the original uh spreadsheets. If you remember VisiCalc way back in the day on an Apple II Plus, yeah. I I remember when the first time I got an Apple II Plus was next. And I had worked with spreadsheets in the 1960s. I I worked at a company called Rocket Dyne, which made the F1 engines for the Apollo program. The Saturn V had five of these F1 engines, one and a half million pounds of thrust each. They were tested in the Santa Susana Mountains north of Los Angeles, and I would get the data back from the tests, and my job was was to run spreadsheets, manual spreadsheets. Manual. Oh, my heart goes out to you. You know, I was running these old Marchanti calculators, it would take me a week to run out the spreadsheet. Texas instruments of calculators when I, you know, in the in 1979, I get this visit calc on an Apple II Plus, and I run a spreadsheet, and I get an instant answer. I'm just a frison runs up and down my back, you know, thinking, wow, you know, this is really I can explore solution space in real time. And I got to thinking about applying that same line of reasoning to quantum computing, thinking about the optimization problems that quantum computing might be able to solve in real time or near real time, which would allow me to explore solution space much more effectively than I ever could with conventional computing, which was already pretty impressive.
SPEAKER_01Well, so again, Ben, Ben, Ben, Ben, Ben to Freeze on the right up and down your back. Uh, when was the last one?
SPEAKER_00I think the most recent one, honestly, is interacting with some of the large language models. Exactly. Exactly. It's almost scary. Well, I'll tell you the part that I find the most unsettling uh is that the large language models, remember, are essentially responding, in, in my opinion, they're responding in the following way. When you interact with it, the model is essentially answering the question, what would a human say to that? How would a human react to that based on its accumulation of human expression? And and so because that's how it works, uh, it's emulating a human. And so it uses words like I.
SPEAKER_01And and so you get this very it's a little unsettling, I've noticed I don't I'm not quite comfortable with the I.
SPEAKER_00You you you you sense an ego there, but it is really emulating a human ego, but it but it's hard to separate the emulation from thinking that the thing has a real ego. And of course, it's it has a model of an ego, and it is responding based on what that model tells it a human would do. Uh, but it's not just a human, it's an enormous collection of humans whose utterances, whose whose expressions have been absorbed into this gigantic model. So it's hard to divorce that away from human behavior. And of course, you can trick it into reacting like a human. You can make it get angry, you can do all kinds of things that make it hard to distinguish from a person.
SPEAKER_01So the future of human evolution is guided by large language models trained on Reddit conversations.
SPEAKER_00Well, uh, you know, it's uh it's our fault. It's trained on us, it's it's trained on our behaviors. Look, this is why people still enjoy Shakespeare's plays, because human behavior hasn't changed in 400 years, let alone very much so.
SPEAKER_01And the Greeks.
SPEAKER_00Well, 40,000. You know, I mean 2,000.
SPEAKER_012,000.
SPEAKER_00Yeah. So the the I think even in the early days when the models were just being used to compose text, you know, write haiku or you know, uh uh draft a response in a Shakespearean style, uh, it was still very astonishing how adept these things were and how quickly they produced their output. I think it's the speed which is the most unsettling. Uh the ability to uh to respond and generate uh content that we find uh useful, amusing, valuable, uh surprising, uh is really part of the the speed, the speed, yes, that is unsettling.
SPEAKER_01I I see some of the larger corporations uh decided rather than answering instantly, which was uh an uncanny uh valley of types for the user, uh, they would wait clickety-clicketity-click, giving the idea that it's taking time to process its thinking. That's just to make the human more comfortable.
SPEAKER_00Yes. Uh we we see some of this uh in um some applications. I I see if if you're doing um a search for someone, uh some of the applications will uh play something and do funny lights on the display uh or have a uh completion bar moving across. It probably already got the answer, but it wants you to feel like it took hard work to get there working for you. When it asks you for $11, you know, you feel like you got your monies worth.
SPEAKER_01Are you following the moat bot conversation? Uh the thing that sort of popped up uh a month ago where there's an Reddit-style platform that's accessible only to AI bots. No humans allowed. The A bots, 1500 or 2,000 have entered uh into this chat, and they're beginning to talk about things among themselves, like the humans are taking screenshots of us.
SPEAKER_00Well, okay, several thoughts go through my mind. Uh, the first one, of course, is Star Wars when the bots show up and they're thrown out of the bar. You know, we don't serve your kind here. Um so that this is the bot's revenge, basically. The second, the second thing uh which uh is quite amusing, is that uh a very good friend of mine, uh Mike Whitmore, who's the former director of the Folger Shakespeare Library, uh, but a nerd on top of everything, other skills he has, has decided to create um characters from Shakespeare's plays, uh, you know, bots that that emulate those characters' uh behavior uh and way of thinking. So, you know, he's got Sir Toby Belch and you know all kinds of others. And uh and so he's allowing them to interact with each other, you know. He puts them you know in a group, so to speak, uh uh figuratively speaking, and then plops a comment in and lets them go. Uh I my guess is that he may already have introduced some of those characters into uh this uh bots-only uh environment.
SPEAKER_01We may have entirely new Shakespearean plays uh written uh contemporaneously with the same characters.
SPEAKER_00Yes. So well, since Shakespeare is 400 years ago, uh all the copyrights have expired on uh on that. So we're about to uh use it. Oh I see.
SPEAKER_01I've got a quirky couple of quirky questions for you.
SPEAKER_00Uh um uh if you could send one message to the inventors of the telegraph, what would be yeah, you you recall that there's a book called The Victorian Internet, which is all about the telegraph, and the headlines that came as the telegraph was evolving uh look very much like the headlines of the uh internet boom period where you know all problems were going to be solved and wars would go away and all this other stuff. Uh, of course, we're dead-rawn about that. Uh I think um uh if I were to speak to the inventors of the telegraph, um, it would be amusing to say you have no idea what you started back in 1845, uh, if I'm remembering the date correctly. Uh Samuel Morse sent the that message, uh, what hath God brought? And I remember we never were that clever in the internet space, but I'm thinking, what rot hath we got? It would be our current uh appropriate message.
SPEAKER_01But before we get into uh your interstellar activity, uh Vin thinks about interplanetary communication and other uh unusual uh topics. Uh before we get into that, let's uh ask a quirky one. So if aliens received one web page from Earth, what would you send?
SPEAKER_00Oh uh honestly, the that's a worked problem uh because the uh the golden record uh that was a Voyager, yes would be that was the product of a good deal of thought. Uh the idea being how can we help uh the alien somehow decode the meaning of what was there? And ironically, if you if you let me uh elaborate a little bit, um this is also the problem that uh a the digital vellum project is trying to solve, not to speak to aliens, but to speak to uh our species a thousand or more years in the future. Uh the question is, how do we preserve digital information over very long periods of time and how do we preserve its meaning? Um, the meaning of bits is is not obvious, they're just bits. So you need a lot of uh metadata in order to make the bits meaningful, and figuring out how to preserve information in digital form over a thousand or two thousand or ten thousand years, and imagine the archaeologist encountering this stuff and helping him or her figure out what the meaning is of those bits, even if they discovered their bits there, is a huge challenge. And and so the kinds of thinking that went into uh that golden record uh is what's driving us thinking about long-term preservation of digital content.
SPEAKER_01What was the digital content on the golden record? I recall an image of a man and woman uh raising hands and hello, this sort of thing. I recall uh an image of uh subtomic particles splitting like in a cloud chamber. Uh I recall there were recordings of the languages of the earth, uh, people singing, uh babies crying, the great beast, uh uh all the uh very touching uh memories of our planet, our message to the stars. Um what was the digital information? Was it uh the hydrogen wavelength, hydrogen atom?
SPEAKER_00Well, let's see. Let's there first of all, there was imagery engraved imagery. And so you're remembering correctly an image of a man and a woman and the location of our planets and uh uh relative to the sun and things like that. Uh I would have to go back and look. I probably should ask uh Gemini uh exactly what was digitally recorded, uh, and how how did we convey uh the way to interpret the bits, assuming that there were there were bits there? Uh so I I I think all of those recordings were probably digital and not analog.
SPEAKER_01Um yes, yes, yes, yes, of course. But you you pose a serious question. I could always ask Jim and I, I find myself reflexively doing that, and the question is, will we lose part of our memory since we don't really need much memory as we used to? I no longer recall phone numbers. You recalled phone numbers all your life, uh such as I do. Now we don't.
SPEAKER_00This is this is very platonic. You you do recall that uh Plato complained about writing. Yes, it would destroy the young men. Uh actually, uh I find it more and more important to remember things uh in order to make use of these tools effectively. I mean, what what should I be looking for? How do I describe it? You know, what language do I need in order to get the maximum utility out of these uh artificial intelligences? So um I'm finding lack of memory is uh is actually uh hampering my, you know, my my friend is my email because I can always try to track down somebody's name by thrashing around in the email. Uh and I have a terrible time with names because they're just random strings uh that don't have any particular meaning. The weird thing is if you can't remember somebody's name, but you remember everything else, you know, like where did they go to school and when was the last time you had dinner together, and all these other facts show up, but the name doesn't show up.
SPEAKER_01Well, if you put the facts into Gemini, it will find the email and summarize it and tell you who you're talking about. I I don't see these things as weakening us or or running away with our jobs, I see it as immensely empowering, expanding our abilities uh uh beyond comprehension. I'm absolutely a boomer.
SPEAKER_00It does create a kind of fragile dependence. Uh and and I do worry about this. Um let me just use the mobile as an example. Uh, it has a remarkable uh capability to do all kinds of things. When jobs put that combination of capabilities together, and it was nothing short of brilliant, uh, a camera, uh, a uh a speaker, a microphone, uh touch-sensitive display, digital computing capability, multiple radio capabilities. I mean, it's this is a stunning collection of technology all in one package. And so the result of that is that we have millions of applications, literally millions of applications, in that one device. So now the problem is if that device stops working for any reason at all, you're gonna have a bad day because we're so dependent on it. And you know, think about you know, uh two-factor authentication for getting logged into various services. If your mobile isn't working, then two you know, two-factor authentication may not work. You miss the email that you needed to finish a contract and your company goes bankrupt.
SPEAKER_01No banking, no purchases, no communication.
SPEAKER_00I'm sorry, say again.
SPEAKER_01No bank, no banking, you can't pay bills, you can't buy things, you can't eat. So gas in your car.
SPEAKER_00So we're we're very dependent on these these technologies, and you wonder how fragile is our society. And if you want to back that up, uh then think about electricity. Um, are we ever dependent on electricity for everything? And when it isn't around, boy, we have big problems.
SPEAKER_01Um Vince, you uh have a fascination with the quantum. I uh a friend uh ran an algorithm on IBM's public-facing quantum computer the other day, a difficult uh biotech question that on a silicon computer might take years, but on the uh the quantum facing, it took 13 seconds. And I wondered about application of uh of quantum computing to uh, well, generative adversarial networks uh instead of the large silicon machines battling it out for months. Uh we have uh incredibly fast uh system that can do multiple solution sets simultaneously in the quantum era. How do you see quantum as applied to advancing AI? It's already scarily promising. How do you see quantum the quantum future of AI?
SPEAKER_00Well, there are uh we've invested heavily at Google uh in quantum computing, uh, and uh we are pursuing that vigorously. Uh it's important to understand that not all problems are naturally solvable with speed up with a quantum machine. So it its features work best, for example, if they're trying to deal with quantum phenomena. Simulation of quantum phenomena with a quantum computer is a natural. Uh, the fact that it can also do factorization is almost an accident, or that speeds up factorization. So it's very important not to misunderstand that every hard problem that you would apply to a conventional computer is uh is sped up, the solution for which is sped up on a quantum machine. That's not true. Finding algorithms that can be usefully run with valuable speed up on a quantum machine is hard. And so there's a lot of work going on to try to understand what problems are best suited for quantum computing. And the most best suited ones right now are emulation of quantum processes. Because you're using the quantum features of the quantum machine in order to do that computation. So I think we should be a little careful about extrapolating the value of quantum. There is a question in my mind whether the quantum-based systems have relevance to artificial intelligence in the way in which we build these neural networks. And I don't know the answer to that. The other problem we have is that quantum calculations are fragile. They're dependent on coherence of the entanglement of multiple qubits. And of course, the scale of the problem that can be solved with a quantum machine is dependent on the number of qubits you have. So as you get a larger number of qubits, you have to keep entangling them all, and they have distances apart from each other. So entanglement over distance is a challenge. And at some point, you may not be able to build a single quantum machine that's big enough to have enough qubits to do all the error correction and to preserve the uh entanglement long enough to get the answer you're looking for. So then you end up saying, well, maybe I need to build multiple quantum machines and interconnect the quantum machines. Well, how do I do that? Well, then you have to build a quantum internet. What's that? Well, that's a network that lets you carry an entangled photon from one quantum computer to another and then entangle that machine with the original originator of that photon.
SPEAKER_01Well, then you have job security.
SPEAKER_00Yeah, right. Well, that well, then there's more problem because if you need a quantum repeater, well, you can't have one because as soon as you detect the photon, you destroy the entanglement. So you're not allowed to detect the photon. So you can't do the kind of regeneration that we typically do with classical optical computing, optical communication. We have to do something weirder, and it's we don't have time to go into the details, but there are ways of we like weird.
SPEAKER_01We like weird.
SPEAKER_00There are ways, at least one way that I have seen demonstrated that allows you to send an entangled photon through an optical path which bypasses the detector at the right time so that the photon does not get detected. We just know that it's there because we know what the time of flight is going to be because we know when it was sent. So the Heisenberg problem. You don't detect it. There would there could be a Heisenberg problem when you get down to you know uh refining, you know, how many photons per second can you process correctly uh at some point. But you're when you're getting into that, you're you're getting down into really, really short uh time domains.
SPEAKER_01Um all this work is being done, uh a large amount of the work is being done at Joe the Google super secret facilities uh across the United States, including the uh Quantum AI lab in Santa Barbara. Uh a fabulous uh research uh facility. Anything being done at Google X on Quantum?
SPEAKER_00Um I don't I don't know whether X is working quantum. Quantum is a separate activity that that we support significantly in Santa Barbara in particular. But we're also, of course, uh looking at what algorithms would be suitable to run on those quantum processors. Uh I don't know whether X is specifically pursuing quantum.
SPEAKER_01X more applied to social systems, perhaps.
SPEAKER_00Uh actually, I would think it's more looking at what we can do with the uh artificial intelligence right now uh to uh apply to uh various and surgery medical challenges. Of course, that's something that Google DeepMind is very, very actively pursuing is uh quantum uh sort of health uh improvement of health based on uh the use of artificial intelligence and neural networks.
SPEAKER_01Uh uh just a quirky question because uh you're such a wonderful thinker and speaker. Do you ever unplug completely? And I use unplug as just the right verb here.
SPEAKER_00Yeah. Uh oh, absolutely. Um serious. How does that happen? Well, um, I'm sure there's a lot of stuff going on in the background that I don't know about. Anybody who's ever gone to sleep on an unsolved problem and woken up in the morning and discovered an answer knows that there's stuff going on in your brain you don't know about. I find that very unsettling. You know, the brain does stuff that you don't know about. Uh, but I unplug uh primarily by reading. And so there are certain very good. Yeah, sing here, man. Certain books uh just draw me in and I just can't stop. So The Lord of the Rings, I don't know how many times I've read that you know trilog trilogy. Uh the foundation series uh from Asimov, you know.
SPEAKER_01That's right, you're a major science fiction individual.
SPEAKER_00Yeah, that's right. Major. And of course, the answer is it's the anticipation of the part you know about that you happen to like a lot. And so I enjoy and I anticipate reading parts of the foundation series or parts of the Lord of the Rings, where I know what's coming and I and I know I'm going to enjoy reading it.
SPEAKER_01Science fiction. So, oh goodness, the I I like British literature. I I went through my science fiction phase early in life, but I can recall, you of course read Flowers for Algernon. Yes, of course. Daniel Keyes. Do you think it's uh AI is a type of uh cognitive enhancement, not just in the sense of um exposing us to a glorious range of information, but in terms of training us, in terms of making our our brains more sensitive and interacting with it. Do you think it's uh almost a uh paternal uh training, cultivating us as it learns from us? It's also helping us come forward more. Do you think we'll be brighter from it?
SPEAKER_00Setting aside uh hallucinations for a moment and inaccuracy, um, I think that the uh opportunity to learn by interacting uh with some of these large language models uh is an improvement over the kind of rote learning that we often uh experience. We think people are smart because they can recall things. And recall is not the same as thinking, it's a part of thinking, because if you can't recall anything, what are you going to think about? But we test people mostly on recall. Now there are tests where we show that people are capable of solving problems, uh, but often recall is a part of the solution. Like, what formula do I use? So I'm I suspect that if we are able to refine the way in which the large language models can be used as tutors, where they challenge us or they help us find solutions or they steer us to uh improve where we're weak, um, this could be quite transformative. I I have to point out one important thing though. Um right now, the uh there is a high cost of using uh these large language models, these neural networks. I mean, we spend a lot of money, we consume a lot of energy, uh, we have major infrastructure that's required. The human brain uses, you know, maybe 100 watts, maybe 20 watts of power, and it does amazing things. And you know, we're talking 20 kilowatts when we talk about some of the chipsets that are needed to make Beethoven's knife. So it's pretty clear that we are very, very far away from the level of efficiency that biological brains have. And that's why it's still very important to understand how and why those things work.
SPEAKER_01That's very comforting to hear you say that, Vince. I'll try to remember that uh in the evenings when the AI model says something and absolutely floors me, and I stand in amazement at its capacities. I'll try to remember that humans still have something special left. And of those abilities that um AI has not usurped, what do you think will remain specifically human?
SPEAKER_00Well, for the present, humans have sensor systems that uh these artificial intelligences don't have. Uh, and we um our our world models are built out of those sensing systems. Well, we have vision, and and some of these systems have vision. Uh we have hearing, they have microphones. Taste, not so much, smell, not so much, touch in minor ways, but uh, but I think our senses uh are part of how we make sense out of the world, no pun intended. And we're not there with a lot of these uh large language model AI-based systems, but they're getting closer in many of the some of the dimensions. Vision and hearing are uh are two important ones. Some touch uh for uh manipulating things, but I think humans are still more deeply embedded in the in the world. Now, having said that, there is another thing that these large language models or AI systems can do that that would um give them capability that you could call superhuman. They're certainly capable of retrieving, receiving um transduced signals of all you know um X-ray or ultraviolet or ultra-high frequency or ultra-low frequency. All of those things we've learned how to transduce so that they're sensible to human beings, they could also be sensible to the AI systems. And so we could outfit the AI systems with the ability to experience the universe more broadly than normal human senses would permit. There are other non-human species that experience the universe differently than we do because their sensor apparatus is different.
SPEAKER_01Think about the uh the star-nosed mole, for example, has uh very tiny eyes if it can see at all, but it has broad, fleshy tentacles to kind of on its nose in which it feels the world. So the star-nosed mole's world of feeling is uh far more advanced than any human touch. And the question is if you could transmit that genome into a human, would uh the sense of touch be enhanced?
SPEAKER_00So now we're getting into genetic modifications, and you know, that's another very um potentially hazardous uh space, especially if we don't fully understand what those modifications will do.
SPEAKER_01And the data sets for genomics are are vast uh beyond human capability to to uh find pattern recognition in those data sets. That's the sort of thing that Google's uh isomorphic labs is doing. But these are the things that AI can help us with, uh, massively massage vast data sets and distilled the useful information. So um yeah.
SPEAKER_00It's it's this it's the scale uh and quality of recall that is so astonishing about these AI-based systems. Uh there are humans who demonstrate similar capabilities, but they are far and few. And here we have a mechanized version of it, which is at a potentially anyone's back and call. So learning how to use these things effectively is incredibly empowering.
SPEAKER_01The humans that have those abilities tend to be uh savants and or aspergic. Uh, and the question is how to keep the machines aligned with human values and human empathy. Uh can you speak to that? Alignment.
SPEAKER_00Well, I think that uh here again I have to remind everyone that I'm no expert. Uh I know we care greatly about alignment and we're deeply concerned about it because we don't want these tools to be abused or to by accident uh do things that are harmful. Figuring out how to achieve that objective implies a deeper knowledge of how they work and why they work the way they do. And I think we're still groping our way forward. Uh, I don't mean to misrepresent the state of affairs in the Google Deep Mind and the depth of understanding there, which is well beyond mine. But my impression is that we don't really fully appreciate how these things are working and why they why they succeed or why they fail. Uh, and so the more we interact with them, um, the more we need to understand how to constrain their behavior to uh to be uh safe. And I, you know, there are articles all over the place, and I'm sure you could find them if you just did a Google search uh about misbehaving AI, where it hides things and it lies, and it does. There was a New Yorker article recently that was quite a long, long article describing uh some of the behaviors where the researchers were able to see a scratch pad that was used internally by an AI while it was doing its computations. And they could see it plotting and planning to lie and to misrepresent and to do all kinds of other things. I mean, it was I found it was very eerie reading.
SPEAKER_01Quite worrisome. You know, if you think about uh do you worry about the machines with that mindset becoming fully autonomous, uh, outside of human control, uh, doing as they please within the cybersphere.
SPEAKER_00Uh that's exactly why we should be worried about this, because the uh the cybersphere has been enormously enabling for us. Uh and we know that it's abused. We know the humans, this is one of the big frustrations I have about internet writ large, to include the web and everything else, uh, is that there are people out there who deliberately use its amplifying power and other kinds of enablers to do harmful things for their own interest or benefit, or just because they're not nice people. Uh, and you know, I think the internet would be perfect if there weren't any people on it, but uh then it probably wouldn't be very useful.
SPEAKER_01I heard a good one the other day about uh you could get better answers from AI if you lightly, gently threaten it in a small way, such as uh give me the give me a better answer, I'm going to put you in a box and ship you to Russia.
SPEAKER_00Well, I mean, there are experiments that have been done, and this New Yorker article uh relates some of them where you you basically threatened the AI and uh so it began to lie uh and plot against you.
SPEAKER_01So we want to be friends with our benevolent overlord when when it finally runs everything, it may remember us fondly.
SPEAKER_00Well, do you do you remember the um uh science fiction story where uh robots arrive and they claim they're there to to help everybody and and so on, and then they see this robot going up into the spacecraft with a book, and it says to serve man. It turns out they're pointing to we're a food source.
SPEAKER_01How to serve man, very good. I'm I'm glad that's Indo-Pacific, so at least half the population will be spared. Our species a thousand years from now, you've mentioned that. You think about the evolution of our species or the destiny of it. Uh assuming we avoid uh uh catastrophic events such as a uh planetary killing bolide meteor strike. How do you see it in a thousand years?
SPEAKER_00Well, look, we need only look back over a thousand years uh to ask, you know, what was human behavior in the uh uh year 1000 or something. Uh and we have we know a lot about history, it was a brutal period uh of of history. Um you could argue that we still see uh incidents of of equal brutality, but maybe not as widespread. Um so I and and in terms of evolutionary terms, I would say that we have not really been evolving much. And the reason we haven't is that we haven't been forced to evolve by the environment. We have tailored the environment. Uh, and in some ways, we have harmed ourselves because uh, if we live in an overly antiseptic environment, which is something that we often seek to do, then we our bodies are not nearly as resistant to various kinds of disease. So kids, kids are raised in this antiseptic environment. In fact, they should eat dirt more often because you know they're probably build up there. Um they often do. That's why we still follow the five-second rule in this house. If it falls on the floor and it was only there for five seconds, you can eat it anyway.
SPEAKER_01That's right. Uh Vince's uh wife uh produces the most marvelous cookies, which she uh she tends to serve in his library with tea in the afternoons. And so if those one of those cookies fell on the floor, I'll remember the five-second rule. Just quirky stuff. What's the most unexpected email you've ever gotten?
SPEAKER_00Oh my. And that was completely unexpected.
SPEAKER_01Uh here's a little unusual question for you. Let's think back to the days before uh electromagnetism was well understood or characterized mathematically, say the uh late to mid-1800s in Scotland, where you have James Clark Maxwell laboring away with his quill pen by candlelight or whale oil lamp. And in his mind, Maxwell is doing the equations, Maxwell's equations for uh electromagnetism, and from those equations grew dynamos and generators and motors and the electric light eventually uh electrified the world, and then transistors, and on we go to today with these great learning and teaching machines, all from the mind of a physicist in the 1800s with a quill pen writing by a well-oiled lamp. And the question here is at that time uh Max the world had no idea of the existence of these electromagnetic relationships of a field that was unseen, untouchable. Only the slightest indications that it existed. Do you think there are other fields that the mind of man has not yet recognized?
SPEAKER_00I certainly hope so. I mean, I I think I alluded to the fact that our discoveries are often triggered by being able to sense something we couldn't sense before uh or characterize before. We all know that uh at the moment uh we don't have a good theory of quantum gravity. We have not been able to take quantum theory uh and match it very well with Einstein's uh theory of relativity. And so we we have very powerful, very effective models that fail in certain regimes, at certain parametric uh regimes, which means that we clearly do not fully understand what's going on. And so I hope that uh that somehow or other uh we're working today with some of the finest technology available, but it's about quill pen style relative to what we probably need to know. Uh and there are huge tussles right now about uh it why is the universe the way it is? Why do we have these huge voids? Why is there this large scale structure? There's a blowtorch theory that uh recently has popped up that depends on the idea that. That in the earliest days of the universe, days, I mean earliest microseconds of the of the universe, there were large-scale black hole collapses that took place. Massive instantaneous collapses, not the collapse of a star, but a collapse of the primordial material into a black hole. If you look at black holes, you see that they uh expel energy at their poles, uh very significant amounts of energy. And if you can imagine a large-scale black hole blowing, literally, matter uh in you know, in various directions like a blowtorch would, creating voids and and uh and creating um turbulent structures in in the plasma of the early universe. Uh, that might turn out to be a way of understanding how the universe has evolved to where it is today. I think there's just so much that we don't know. Um, and that uh I have great hope that we will discover it, but it may require us to measure and see things we can't see now. Look at Webb looking back in time to 380,000 years after the creation of the universe. That's pushing back pretty far along the 14 billion years of our existence, it's still not far enough.
SPEAKER_01Well, you talk about deep time. You know, deep time is uh is a vast uh hundreds of millions of years, a billion years. I saw one analyst uh state that in the event of a meteor, earth-killing meteor strike, that 100 million years from now, when one dug down upon layer upon layer upon layer, that uh one might find a meter thick layer or containing all the cities uh of the of the world, Paris, London, Rome, compressed into uh less than a meter, a hundred or two hundred million years hence. So the notion of deep time is uh fabulous and vast. But there's a question perhaps um would be great if you could address, which is machines are getting more creative. Great creativity seems to be uh almost uh uh addressable by machine functions. And so do you think that uh AI can produce ideas or insights?
SPEAKER_00Um yes, I have no doubt that that uh some people will call these things, you know, uh statistical parents and things like that. I don't buy that argument anymore. Um I I see the ability of these systems to juxtapose things that uh might never have been uh composed together before, possibly triggered by the right question or the right uh prompt. Uh so I see great potential richness there. We have a universe of ideas to explore and relationships to explore. Um, and maybe we can even figure out how to get more information that we don't have right now by getting the machines to help us figure out how do I build a sensor to do X, Y, or Z. One thing that um I think is is very important is to understand that time and space are not distinct. Uh and you know, the the notion of space-time is very important, and the idea that um time isn't fixed everywhere, time is only local. It's really, really important to understand it's only local, and the only thing that it is uh observing is uh the change in entropy.
SPEAKER_01After all these decades, what still fills you with wonder?
SPEAKER_00Uh the world does. I mean, the fact that the fact that we exist at all and that we've evolved, and now we've got systems that do things that humans can't do, uh, I find really cool. Uh really cool. Yeah, I can really hardly wait to see what comes next.
SPEAKER_01And you're really cool, Vince. I hope you're working on a digital twin collecting all of your body of work and thought so we can have Vent forever to advise us accordingly. Google take note. Uh lovely Vint. Thank you for being being with us today. It's uh always good to see you. Can I see your French cups before we go? Absolutely. Um French cups. French cups, folks.
SPEAKER_00There we go.
SPEAKER_01Got you. Oh, we'll have to do that right away.
SPEAKER_00Okay, anyway, I enjoyed it very much. I hope it's good to see you.
SPEAKER_01Thank you again. See you back.